Datasetpublic

jensjepsen/danish-textman-v2

danish-textman-v2 Danish text-manipulation tasks over Wikipedia passages. Supersedes jensjepsen/danish-textman-v1, which is identical except that it also contained a textman_extraction subtype. split rows train 99,610 validation 4,997 subtype rows textman_summary 20,937 textman_genre_transform 20,934 textman_style_transfer 20,923 textman_rewrite 20,912 textman_elaborate 20,901 What changed textman_extraction is removed. It… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-textman-v2.

sourceHugging Faceapache-2.0updated 22d agoView on Hugging Face

0likes77downloads

Dataset Card

danish-textman-v2

Danish text-manipulation tasks over Wikipedia passages. Supersedes `jensjepsen/danish-textman-v1`, which is identical except that it also contained a textman_extraction subtype.

split	rows
`train`	99,610
`validation`	4,997

subtype	rows
`textman_summary`	20,937
`textman_genre_transform`	20,934
`textman_style_transfer`	20,923
`textman_rewrite`	20,912
`textman_elaborate`	20,901

What changed

textman_extraction is removed. It asked for entities as JSON under ONE fixed schema -- people / places / dates / numbers -- across all of its rows, so it taught a single hardcoded key set rather than the ability to extract whatever keys a prompt asks for. 26% of its numbers values did not appear in the source passage.

Structured extraction now lives in `jensjepsen/danish-extraction-v1`, where the schema is proposed per passage, every value is checked to be a verbatim span, and absent fields are real abstention targets.

The five remaining subtypes are byte-identical to v1, so results comparing a model trained on v1 with one trained on v2 isolate the removal.

Known limitation

87% of passages carry Wikipedia navigation cruft (Referencer, Eksterne henvisninger, Se også) inherited from the source dump, and 1.2% of instruction tails contain a non-Danish word. Neither is addressed here.